Harnessing GPU Power for Enhanced OLTP: A Study in Concurrency Control Schemes

TL;DR

This paper explores the design, implementation, and evaluation of concurrency control schemes on GPUs to enhance OLTP performance, filling a research gap and providing a foundation for future GPU-accelerated transaction systems.

Contribution

It introduces the first GPU-based test-bed for concurrency control schemes in OLTP, implementing and evaluating eight schemes to guide future GPU-accelerated OLTP system development.

Findings

GPU concurrency control schemes can significantly improve OLTP throughput.

Evaluation reveals differences in scheme performance on GPUs versus CPUs.

The test-bed aids in designing high-performance GPU-accelerated OLTP systems.

Abstract

GPUs, whose performance has gone through a huge leap over the past decade, have proved their ability to accelerate Online Analytical Processing (OLAP) operations. On the other hand, there is still a huge gap in the field of GPU-accelerated Online Transaction Processing (OLTP) operations since it was generally believed that GPUswere not suitable for OLTP in the past. However, the massive parallelism and high memory bandwidth give GPUs the potential to process thousands of transactions concurrently. Among the components of OLTP systems, Concurrency Control (CC) schemes have a great impact on the performance of transaction processing and they may behave differently on GPUs because of the different hardware architectures between GPUs and CPUs. In this paper, we design and build the first test-bed gCCTB for CCschemes on GPUsandimplement eight CC schemes for gCCTB. These schemes include six common schemes previously designed for CPUs and two schemes designed for GPUs. Then we make a comprehensive evaluation of these CC schemes with YCSB and TPC-C benchmarks and a number of launch parameters on GPUs. The experience accumulated on our test-bed can assist researchers andengineers to design andimplementnewGPU-acceleratedOLTP systems. Furthermore, the results of our evaluation cast light on research directions of high performance CC schemes on GPUs.